GMM: Greedy Matching & Merging in Brachytherapy

• GMM is a two-stage reconstruction method that first greedily pairs CT-detected needle tips and handles, then merges duplicates to build accurate 3D needle trajectories.
• The approach formulates the pairing as an unbalanced assignment problem with constraints on length, angle, and non-crossing conditions to ensure valid matches.
• Empirical evaluations indicate improvements in precision and F1 scores over segmentation-based techniques, highlighting its efficiency in clinical needle localization.

Greedy Matching and Merging (GMM) most literally denotes the constrained tip-handle association procedure introduced for multi-needle localization in pelvic seed implant brachytherapy, where detected needle tips and handles are greedily paired and duplicate paths are subsequently merged to reconstruct 3D needles (Xiao et al., 22 Sep 2025). In a broader research sense, however, the label is not a standard name for a single established framework. Several related papers study greedy matching combined with contraction, queue-aware joint actions, or post-matching grouping, but explicitly do not present a method literally called GMM. This terminological split is central: the 2025 brachytherapy paper is the literal source for the name, whereas earlier work provides adjacent algorithmic ideas rather than a canonical GMM formalism (Dietzfelbinger et al., 2012).

1. Terminology and conceptual scope

In the literal sense, GMM is a second-stage reconstruction algorithm inside a two-stage pipeline. The first stage detects needle tips and handles from axial CT slices; the second stage solves a constrained tip-handle association problem and then merges duplicate reconstructions (Xiao et al., 22 Sep 2025). In the broader literature, “matching” and “merging” appear in several nearby but non-identical forms: safe contractions in graph matching, multi-item joint matching in queueing systems, and grouping matched pairs into larger clusters. These are close analogues, not terminological equivalents.

Source	Core mechanism	Relation to GMM
(Xiao et al., 22 Sep 2025)	Greedy tip-handle pairing plus duplicate-path merging	Literal GMM
(Dietzfelbinger et al., 2012)	Degree-2 contraction in greedy maximum matching	Merging-like safe contraction
(Nazari et al., 2016)	EGPD matching in a virtual system with negative queues	Queue-aware analogue
(Anshelevich et al., 2015)	Greedy matching used as a black box for clustering	Matching-to-grouping reduction
(Blelloch et al., 2012)	Parallel greedy maximal matching under random order	Greedy matching only
(Eden et al., 2018)	Ordering-based greedy bipartite matching; proof-side path merging	Not operational GMM
(Adamczyk, 2010, Deligkas et al., 2016)	Greedy stochastic matching; weighted greedy matching complexity	Greedy matching foundations

This suggests that GMM is best treated as a family resemblance across several strands of research rather than as a universally standardized algorithmic class. The strictest usage is the clinical reconstruction method of (Xiao et al., 22 Sep 2025); broader usage is interpretive.

2. Formal problem formulation

The brachytherapy formulation begins with detected needle tips and handles after slice-wise 2D detection and 3D fusion. The detector is an anchor-free network based on HRNet that predicts, for needle tips and handles, a heatmap indicating center probability, a local offset to refine the center location beyond heatmap quantization, and a polar angle indicating the 2D orientation of the needle at that endpoint. Detection outputs across slices are combined to produce 3D positions $(x_o, y_o, z_o)$ and orientation $\alpha_o$ for candidate tips and handles (Xiao et al., 22 Sep 2025).

The association stage is formulated as an unbalanced assignment problem with constraints (UAP-C). The detected tips are $T=\{t_1,t_2,\dots,t_m\}$, the detected handles are $H=\{h_1,h_2,\dots,h_n\}$, and the binary decision variable $x_{i,j}\in\{0,1\}$ indicates whether tip $i$ is matched with handle $j$. The objective is

$\max \sum_{i=1}^m \sum_{j=1}^n o_{i,j} x_{i,j},$

subject to one-to-one partial matching,

$$\sum_{i\in T} x_{i,j}\leq 1 \quad \forall j\in H,\qquad \sum_{j\in H} x_{i,j}\leq 1 \quad \forall i\in T,$$

length consistency,

$$|L_{i,j}-L_{\text{prior}}| < \varepsilon_l \quad \forall i\in T, j\in H,$$

angle consistency,

$\alpha_o$0

a non-crossing condition $\alpha_o$1, binary feasibility, and the known-cardinality constraint

$\alpha_o$2

The scoring function is defined from CT intensity statistics sampled along the candidate line segment joining tip $\alpha_o$3 and handle $\alpha_o$4: $\alpha_o$5 where $\alpha_o$6 is the average Hounsfield Unit value and $\alpha_o$7 is the corresponding standard deviation. The paper uses $\alpha_o$8 in the optimization statement and then writes the score as $\alpha_o$9; these refer to the same pairing criterion. The 3D Euclidean distance

$T=\{t_1,t_2,\dots,t_m\}$0

is the geometric length used in the prior-length constraint. The non-crossing constraint is checked using the shortest distance between two line segments (Xiao et al., 22 Sep 2025).

3. Greedy matching and merging procedure

The algorithm first forms all possible tip-handle pairs from the 3D candidate sets. For each pair it computes the Euclidean distance $T=\{t_1,t_2,\dots,t_m\}$1, the angle difference $T=\{t_1,t_2,\dots,t_m\}$2, and the path score $T=\{t_1,t_2,\dots,t_m\}$3. These values populate a score matrix $T=\{t_1,t_2,\dots,t_m\}$4. Pairs that violate the length or angle constraints receive the special value

$T=\{t_1,t_2,\dots,t_m\}$5

which removes them from consideration as feasible assignments (Xiao et al., 22 Sep 2025).

The greedy stage then repeatedly selects the currently highest-scoring feasible pair. At each iteration, the needle path with the highest score that does not intersect with any previously selected paths is chosen. Accepting a pair also enforces one-to-one matching, so the corresponding tip and handle are removed from further consideration. Because this is a sequential best-first rule, earlier choices constrain later ones. The method is therefore greedy in the precise sense that it makes locally optimal decisions without solving the full constrained combinatorial problem globally (Xiao et al., 22 Sep 2025).

The merging stage is activated when the number of matched paths $T=\{t_1,t_2,\dots,t_m\}$6 exceeds the known number of implanted needles $T=\{t_1,t_2,\dots,t_m\}$7. This overproduction is interpreted as duplicate reconstructions. The procedure first targets needle paths with lower score values. Two reconstructed paths are treated as duplicates when their tips are within $T=\{t_1,t_2,\dots,t_m\}$8 and their handles are also within $T=\{t_1,t_2,\dots,t_m\}$9. Their tip positions are then merged, and their handle positions are merged, using weighted sums whose weights are the average non-zero HU values within the corresponding detected mask regions. A faithful formalization of the endpoint update is

$H=\{h_1,h_2,\dots,h_n\}$0

with the same construction applied to handles. Merging is repeated until the number of remaining needle paths equals $H=\{h_1,h_2,\dots,h_n\}$1 (Xiao et al., 22 Sep 2025).

This division into greedy selection and endpoint-level merging is what gives the method its name. Matching resolves the unbalanced assignment under geometric and physical constraints; merging reconciles duplicates caused by slice-wise detection ambiguity and enforces the known final needle count.

4. Empirical behavior and implementation characteristics

The literal GMM paper evaluates the full detector-plus-GMM pipeline on a dataset of 100 patients and compares it with a segmentation-based method utilizing the nnUNet model. The reported 3D needle localization results are: recall $H=\{h_1,h_2,\dots,h_n\}$2 versus $H=\{h_1,h_2,\dots,h_n\}$3, precision $H=\{h_1,h_2,\dots,h_n\}$4 versus $H=\{h_1,h_2,\dots,h_n\}$5, and F1 $H=\{h_1,h_2,\dots,h_n\}$6 versus $H=\{h_1,h_2,\dots,h_n\}$7, with statistically significant improvements in precision $H=\{h_1,h_2,\dots,h_n\}$8 and F1 $H=\{h_1,h_2,\dots,h_n\}$9. For correctly detected needles, the tip 3D localization MAE is $x_{i,j}\in\{0,1\}$0 mm for the segmentation-based method and $x_{i,j}\in\{0,1\}$1 mm for the proposed method, with $x_{i,j}\in\{0,1\}$2. Handle localization and 3D angle estimation are described as comparable between methods (Xiao et al., 22 Sep 2025).

The implementation details matter because the matching score depends on image intensities along candidate line segments. Before matching, CT slices are enhanced using a white top-hat transform applied slice by slice with a spherical structuring element radius of 5 pixels. Raw HU values are thresholded at 800 HU and then linearly normalized to $x_{i,j}\in\{0,1\}$3. Candidate endpoint detections are obtained by searching local maxima in the 8-neighborhood of each heatmap pixel and retaining local maxima above a predefined threshold, although the numerical threshold is not reported (Xiao et al., 22 Sep 2025).

The paper reports a time complexity of

$x_{i,j}\in\{0,1\}$4

for GMM, where $x_{i,j}\in\{0,1\}$5 is the number of detected tips or handles. The matching time is $x_{i,j}\in\{0,1\}$6 s per case, detector inference time is $x_{i,j}\in\{0,1\}$7 s per case, and total pipeline time is $x_{i,j}\in\{0,1\}$8 s per case. In the largest tested case, with 44 needles, matching took 24 s; in a small case with 5 matched needles, it took 0.34 s. The paper also states that there is no dedicated ablation isolating greedy matching alone, merging alone, or matching against Hungarian or integer-programming alternatives, so the evidence for GMM is indirect and comes from the full pipeline rather than from a componentwise study (Xiao et al., 22 Sep 2025).

5. Relations to adjacent greedy matching literatures

The closest graph-theoretic analogue to GMM’s “merging” is the degree-2 contraction used in sparse-random-graph maximum matching. In that setting, if a vertex $x_{i,j}\in\{0,1\}$9 has degree 2 with neighbors $i$0, the algorithm contracts $i$1 into a supernode via

$i$2

solves recursively, and then lifts the solution back. The recommended heuristic, OPT(1,2):HEU(pot,deg), combines degree-1 reduction, degree-2 contraction, and the potential score

$i$3

Experimentally, it had zero failures on all tested bipartite graphs and only three failures out of all 100 instances per $i$4 on general graphs, occurring at $i$5, each with failure rate $i$6. This is not called GMM in that paper, but it is the clearest example of a safe contraction acting as a merging-like reduction inside greedy matching (Dietzfelbinger et al., 2012).

A systems-oriented analogue appears in dynamic matching with the extended greedy primal-dual algorithm. There, a matching $i$7 consumes a resource vector $i$8, the virtual queue can become negative, and the control rule is

$i$9

The method is greedy because it makes myopic per-slot choices from the current state, but it is also primal-dual because queue lengths act as shadow prices. The paper does not discuss a separate operation named merge, yet each matching may consume multiple item types jointly, which is explicitly described as a reasonable merge-like interpretation (Nazari et al., 2016).

A third nearby strand uses matching as a black box for grouping. In ordinal matching and clustering, greedy matching picks undominated edges, gives a deterministic ordinal $j$0-approximation, and under the metric inequality becomes a component of a randomized $j$1-approximation algorithm. The same paper derives a $j$2-approximation for Max $j$3-Sum by first finding a matching and then forming equal-sized clusters from matched edges. This is the paper’s clearest matching-to-merging analogue: pairs found by matching become building blocks for larger groups (Anshelevich et al., 2015).

Other papers delimit the boundary of the concept. Random-order greedy maximal matching can be parallelized while returning exactly the same result as the sequential greedy algorithm, with $j$4 rounds w.h.p. and an $j$5-work, $j$6-depth implementation; this is highly relevant to greedy matching but contains no merging stage (Blelloch et al., 2012). Ordering-based greedy bipartite matching proves that there exists a polytime algorithm to compute a permutation $j$7 for which for every adversarial $j$8 at least $j$9 fraction of the vertices of $\max \sum_{i=1}^m \sum_{j=1}^n o_{i,j} x_{i,j},$0 are matched, and also proves that certain high-degree regular graphs satisfy the upper bound $\max \sum_{i=1}^m \sum_{j=1}^n o_{i,j} x_{i,j},$1; however, its “path merging” is only a proof-side operation on a spoiling graph, not an operational merge rule in the matching process (Eden et al., 2018). In stochastic matching, the simple policy that probes an available edge with maximum success probability is a $\max \sum_{i=1}^m \sum_{j=1}^n o_{i,j} x_{i,j},$2-approximation, and successful probes remove their endpoints from the residual instance; this offers a recursive residual-update template rather than an explicit merge operator (Adamczyk, 2010). In weighted greedy matching, computing a maximum-weight greedy matching is strongly NP-hard and APX-complete, even on graphs with maximum degree at most 3 and with at most three different integer edge weights, although the problem becomes polynomial-time solvable when $\max \sum_{i=1}^m \sum_{j=1}^n o_{i,j} x_{i,j},$3 and trivial when $\max \sum_{i=1}^m \sum_{j=1}^n o_{i,j} x_{i,j},$4; this sharpens the distinction between generating a greedy matching and optimizing over all greedy matchings (Deligkas et al., 2016).

6. Theoretical status, misconceptions, and limitations

A common misconception is that GMM names a mature, unified framework with a shared formal theory. The evidence does not support that reading. The literal GMM method is a domain-specific heuristic for constrained endpoint association in pelvic CT; it does not guarantee optimality, and the authors explicitly state that obtaining a global optimum in ideal time is difficult. They propose future work involving perturbation operators such as variable neighborhood search, which confirms that the current method is a balance between computational efficiency and solution quality rather than an exact solver (Xiao et al., 22 Sep 2025).

A second misconception is that “merging” has a stable meaning across the literature. In related papers, it may mean safe contraction of a degree-2 structure, grouping matched pairs into clusters, a queue-aware joint consumption action, or merely a proof-side transformation. Several papers explicitly say that they do not define or use a method called GMM. This suggests that the phrase is best handled with terminological care: literal when referring to the brachytherapy method, analogical when referring to earlier greedy matching literatures (Dietzfelbinger et al., 2012).

The theoretical landscape is correspondingly heterogeneous. Some neighboring formulations have strong guarantees: EGPD is asymptotically optimal as $\max \sum_{i=1}^m \sum_{j=1}^n o_{i,j} x_{i,j},$5; ordinal matching achieves a $\max \sum_{i=1}^m \sum_{j=1}^n o_{i,j} x_{i,j},$6-approximation under the metric inequality; simple greedy stochastic matching is a $\max \sum_{i=1}^m \sum_{j=1}^n o_{i,j} x_{i,j},$7-approximation; ordering-based greedy bipartite matching guarantees a fraction $\max \sum_{i=1}^m \sum_{j=1}^n o_{i,j} x_{i,j},$8; and weighted GreedyMatching is strongly NP-hard and APX-complete. None of these results transfers directly to the literal GMM formulation in pelvic seed implant brachytherapy, because that problem couples image-derived scores, prior needle length, prior needle count, angle consistency, and non-crossing constraints in a domain-specific UAP-C [(Nazari et al., 2016); (Anshelevich et al., 2015); (Adamczyk, 2010); (Eden et al., 2018); (Deligkas et al., 2016)].

The most defensible synthesis is therefore narrow. Greedy Matching and Merging is, in the strict sense, a constrained greedy reconstruction heuristic for pairing detected endpoints and merging duplicates into final 3D needle paths. In the broader sense, it is a useful editorial umbrella for a recurring design pattern: greedy local matching augmented by a structure-reducing or group-forming operation. This broader usage is informative, but it remains an interpretation rather than a standardized algorithmic doctrine.